fix oom

Signed-off-by: Dongfeng Yu <[email protected]>

Author: dongfengy

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -3975,7 +3975,7 @@ def test_eagle3(self, moe_backend, one_model, overlap_scheduler, mocker):
         # https://nvbugs/5590408: 2-Model overlap scheduling has accuracy issue
         pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
                               cuda_graph_config=CudaGraphConfig())
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6,
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4,
                                         dtype="auto")
 
         eagle_model_dir = f"{llm_models_root()}/gpt_oss/gpt-oss-120b-Eagle3"